mini-cognit-cli 0.2.2

The AI coding agent you can read in an afternoon. ~2000 lines of Python, fully hackable.

mini-cognit-cli

The AI coding agent you can read in an afternoon.

pip install mini-cognit-cli

---

Why mini-cognit-cli?

Every AI coding agent (Claude Code, aider, Kimi CLI, Codex) is a black box with tens of thousands of lines of code. mini-cognit-cli takes the opposite approach:

• ~2000 lines of Python — read the entire codebase, understand every decision
• Add a custom tool in ~30 lines — no plugin system, no abstractions, just a function
• Any OpenAI-compatible API — OpenAI, Anthropic (via proxy), Ollama, LM Studio, anything
• Fork and make it yours — designed to be the starting point for your agent

	Claude Code	aider	Kimi CLI	mini-cognit-cli
Codebase	~100k+ lines	~50k+ lines	~30k+ lines	~2k lines
Time to understand	Weeks	Days	Days	An afternoon
Add a custom tool	Plugin system	Moderate	Plugin / MCP	~30 lines
Provider lock-in	Anthropic	Any	Multi-provider	Any OpenAI-compatible
Language	TypeScript	Python	Python	Python

Quickstart

# Install
pip install "mini-cognit-cli[cli]"

# Configure (pick one)
export OPENAI_API_KEY="sk-..."                    # env var
# OR: create cognit.toml with provider config     # config file

# Go
cognit chat

That's it. The agent can read/write files, run shell commands, search code, browse the web, and more.

Features

Tools — file read/write/edit, shell execution, grep, glob, web search, URL fetch, image reading

Session management — auto-save, --continue to resume, --session <id> to pick a specific session

Image support — /paste-image from clipboard, /img <path> to attach, vision models analyze images

@ file references — type @filename with Tab autocomplete to reference files in your message

Agent Skills — drop a SKILL.md in .cognit/skills/ to teach the agent new workflows (compatible with Claude/Codex skill format)

Context compaction — auto-summarizes when approaching token limits, or /compact to manually compress

Extended thinking — --thinking flag for reasoning models

Tool approval — shell commands need confirmation; --yolo to auto-approve everything

Usage

# Basic
cognit chat

# Specify model
cognit chat -m gpt-4o

# Resume last session
cognit chat --continue

# YOLO mode (auto-approve all tools)
cognit chat --yolo

# Extended thinking
cognit chat --thinking

# List saved sessions
cognit sessions

In-session commands

Command	Description
/help	Show all commands
/clear	Clear conversation and start new session
/compact	Manually compress context
/sessions	List saved sessions
/paste-image	Paste image from clipboard
/img <path>	Attach an image file
@path/to/file	Reference a file (Tab to autocomplete)
/exit	Quit (auto-saves session)

Configuration

Create cognit.toml in your project directory or ~/.config/cognit/:

[providers.my_provider]
type = "openai"
base_url = "https://api.openai.com/v1"
api_key = "sk-..."

[models.default]
provider = "my_provider"
model = "gpt-4o"
max_context_size = 128000

Priority: CLI flags > cognit.toml > environment variables.

SDK usage

from cognit import CognitAgent

agent = CognitAgent(model="gpt-4o", api_key="sk-...")

# Add custom tools
@agent.tool(name="my_tool", description="Does something")
async def my_tool(query: str) -> str:
    return f"result for {query}"

# Chat (context preserved between calls)
result = await agent.chat("What files are here?")
result = await agent.chat("Read the first one.")

# With image
result = await agent.chat("Describe this", images=["data:image/png;base64,..."])

Architecture

src/cognit/          (~2000 lines total)
├── cli.py              ← CLI entry point (Typer)
├── app.py              ← REPL orchestrator
├── config.py           ← TOML config loading
├── sdk.py              ← Programmatic SDK interface
├── llm/
│   ├── provider.py     ← Abstract LLM interface
│   ├── openai_provider.py  ← OpenAI-compatible implementation
│   ├── generate.py     ← LLM call + streaming assembly
│   └── message.py      ← Message model (text, images, tool calls)
├── soul/
│   ├── agent.py        ← Core agent loop (the brain)
│   ├── context.py      ← Conversation history
│   ├── compaction.py   ← Context summarization
│   ├── session.py      ← Session persistence
│   ├── skills.py       ← Agent Skills (SKILL.md) loading
│   └── toolset.py      ← Tool registry + dispatch
├── tools/
│   ├── file_read.py    ← Read files
│   ├── file_write.py   ← Write/edit files
│   ├── glob_tool.py    ← File pattern matching
│   ├── grep.py         ← Search file contents
│   ├── shell.py        ← Shell execution
│   ├── web_search.py   ← Web search (DuckDuckGo)
│   ├── fetch_url.py    ← Fetch web pages
│   └── media_read.py   ← Read images for vision
└── ui/
    └── terminal.py     ← Terminal I/O (prompt-toolkit + rich)

Each module has a clear responsibility. The entire flow:

User input → Agent.run() → LLM call → Tool execution → Loop until done → Response

Adding a custom tool

# src/cognit/tools/my_tool.py
from cognit.soul.toolset import Toolset

def register(toolset: Toolset) -> None:
    @toolset.tool(name="my_tool", description="Does something useful.")
    async def my_tool(query: str, limit: int = 10) -> str:
        return f"Result: {query} (limit={limit})"

Register in sdk.py's _register_builtin_tools():

from cognit.tools import my_tool
my_tool.register(toolset)

That's it. The decorator auto-infers JSON Schema from type hints.

Agent Skills

Drop a SKILL.md file in .cognit/skills/your-skill/SKILL.md:

---
name: code-review
---
When asked to review code, follow these steps:
1. Read the changed files
2. Check for bugs, security issues, and style problems
3. Suggest improvements with code examples

Skills are auto-discovered from .cognit/skills/, ~/.cognit/skills/, and compatible with .claude/skills/ and .codex/skills/.

Requirements

• Python >= 3.12
• Works with any OpenAI-compatible API (OpenAI, Anthropic via proxy, Ollama, LM Studio, etc.)

License

MIT